For a long time, classic Retrieval-Augmented Generation (RAG) was sold to businesses as a "silver bullet": dump your documents into a database, hook up an LLM, and your employees can go on permanent vacation. In reality, it turned out that a smart search bar is still just a search bar. If a Tier-2 support engineer doesn't know exactly what to look for in terabytes of logs and Jira tickets, semantic search won't save them. We have hit a wall with tools that can answer questions but cannot ask them. The situation at MTS Web Services (MWS), described by SRE Lead Vladimir Drobot, perfectly illustrates this ceiling. The company built a RAG system based on Jira and Confluence but quickly tripped over the human factor: engineers didn't know how to frame queries during incident post-mortems where the root cause wasn't immediately obvious.
The problem is that in complex B2B processes, knowledge is scattered chaotically. To close a ticket, one must correlate documentation, bug history, and raw traffic dumps. As Vladimir Drobot notes, the main barrier was that for RAG to provide value, the engineer had to realize in advance what data they needed. Consequently, a tool designed to accelerate work was used only sporadically, and its impact on business metrics lingered in the "gray zone" of subjective feedback. The attempt to automate intelligence via search turned into yet another reference manual gathering dust on a shelf because someone forgot to include a table of contents.
From Reference Manual to Digital Employee
The logic of evolution led the MWS team to an agentic architecture. This marks a shift from the "ask and I'll answer" model to a "here is a task—go figure it out" concept. The difference is fundamental: instead of waiting for a query, the system initiates an investigation the moment a ticket appears. It ingests the task description, extracts data from attachments—ranging from PDFs to tcpdumps—and launches a multi-stage analysis cycle. Essentially, this is an attempt to transform AI from a secretary into an autonomous intern who digs through archives without pestering a senior colleague every five minutes.
Technically, this is far more complex than standard prompt engineering. The team utilizes a combination of Prompt Chaining and Parallelization patterns, where the neural network (in this case, Qwen-32B or similar models) decides at each step whether it has enough facts or needs to "walk" over to the database for more info.
Our solution is closer to a workflow because the primary stages of ticket analysis are known in advance.
This statement by Vladimir Drobot marks a significant shift: AI is finally allowed to do more than just hallucinate eloquently; it is being given control over development tools. As a result, instead of just getting a link to a Confluence article, a Jira engineer receives a ready-made subtask featuring a hypothesis, database exports, and a list of similar past outages.
Economics and the Barriers to Autonomy
Agentic workflows are not just a technological upgrade; they are a matter of survival for tech support economics. When a system takes over primary analysis, it directly saves the most expensive resource: Tier-2 engineers' time. However, there is a catch: an agent is useless without high-quality "fuel." In the MTS case, the system's success hinges on how input data is processed. Using specialized models like Kimi K2 to parse heavy documents is an admission that standard LLMs still struggle with real-world corporate chaos.
Integration with internal REST services turns support into a conveyor belt where humans merely verify the output. But to prevent this conveyor from multiplying errors, a high culture of documentation is required. If Confluence is empty and logs are unstructured, an agent simply scales the mess. MTS deployed the solution on its own infrastructure via MWS GPT, which addresses security concerns but leaves the question of scalability open for less mature IT teams. Without a coherent incident history, any "autonomous analyst" remains an expensive toy. The main question now isn't which model to choose, but whether companies are ready to clean up the processes they have ignored for decades.